You are on page 1of 6

Constrained Predictive Control

Of A Servo-Driven Tracking Turret


P. Martin* N. Brignall** M. MacDonald** M.J. Grimble***


*Ricardo, Control and Electronics Department, Shoreham-by-Sea, England, UK
(e-mail: peter.martin@ricardo.com)
**Selex Sensors & Airborne Systems, Avionics Division, Edinburgh, Scotland,UK
*** Industrial Control Centre, University of Strathclyde, Glasgow, Scotland,UK
(e-mail: m.grimble@eee.strath.ac.uk)
Abstract: Vehicle-mounted 2-axis turrets are widely used in high bandwidth tracking systems, frequently
encountered in air-to-ground, ground-to-air and air-to-air targeting. Existing controllers for these systems
are generally implemented in classical proportional-integral-derivative (PID) form. The objective of this
paper is to examine the novel application of constrained model predictive control (MPC) to a Selex turret
simulation. The characteristics of the control problem are well matched to MPC, as hard saturation
constraints are present in the electrical subsystem and a reference trajectory can be generated for several
seconds in advance due to the predictability of a missile trajectory. The state-space model and Kalman
filter are described, and simulation results are presented to demonstrate the validity and superior
performance of the MPC method.

1. INTRODUCTION
Electromechanical systems for tracking from a moving
platform are widely used in defence applications for target
following on board ships, aircraft or ground vehicles. The
targeting system in this paper consists of three main bodies:
an aircraft, the turret azimuth body or forks, and the turret
elevation body or drum. The forks and drum constitute the
outer axes gimbals, which are orthogonal and enable coarse
target following.
The two main bodies, turret azimuth and elevation, are
described by differential equations relating input and
disturbance torques and friction to output angular
acceleration and rate, as in Masten (1996). The aircraft and
forks are linked by a geared motor that rotates the forks about
the azimuth axis. The forks and drum are then linked by a
geared motor that rotates the drum about the elevation axis.
The geared motors and pulse width modulated (PWM) supply
are described by direct current (DC) motor equations and a
proportional-integral (PI) current loop controller with
cascaded delay and scaling. This completes the plant model
for a single axis of motion.
In the Simulink simulation used to generate results later, a
quaternion representation, as in Robinson (1958) and
Mitchell & Rogers (1968), is used to model the three-
dimensional dynamics of the overall system. However, for
control design purposes, it is reasonable to assume that the
azimuth and elevation bodies are independent of one another.
Reaction torque and aircraft motions are then modelled as
disturbances on the single axis descriptions.
This paper presents an investigation into the use of
constrained predictive control for tracking of angle reference
signals. Rate control is not considered, as proportional
feedback with the addition of a backlash filter is of
satisfactory performance. Standard existing angle control is
accomplished using a PI loop with anti-windup and a switch
for reducing rate demand near bottom dead centre, otherwise
known as the nadir.
One disadvantage of classical PI control lies in the fact that
there are several saturation characteristics and hard
constraints that are not explicitly considered in a PI controller
design. When the turret approaches the nadir, the elevation
angle approaches 90 and azimuth rate demand becomes very
large. Predictive control with constraints is very suitable for
this application, as the saturation and hard constraints are
incorporated into the control algorithm. Additionally, it is
possible to predict the trajectory of the target with some
accuracy over a short period. This reference signal can be
incorporated into model predictive control (MPC) to predict
if the nadir is likely to be approached.
Fig. 1 depicts the system block diagram for the azimuth axis.
The elevation axis is almost identical, although r is replaced
by q and by . The various gains and inertia terms are also
different for the two axes. A resolver-to-digital-converter,
H
rdc
, is used for measuring gimbal angle and the rate is
computed using a filtered digital differentiator, H
rate
.
The aim is to control using r
dem
as the input and
tgt
as the
reference signal. Previously, the block marked Angle
Control contained PI with anti-windup and a switch for
reducing rate demand near the nadir, but is replaced by a
predictive controller.

1/s(N
2
J
m
+J
L
) N K
t
1/(R+sL) PI
K
b
N
H
pwm
A/V H
bl
P
rate 1/s
f(r)
H
rate
H
rdc
Angle
Control

r
r
AC
r
AZ
T
m
I
m
V
m
I
ref
I
dem
u
r
dem
r
comp
rdc

tgt

( )
rct AC AC AC
T r q p g , , ,

_ _ _ _
+ +
Fig. 1 System Block Diagram
2. SYSTEM DESCRIPTION
Each axis is described in state-space:
( ) ( ) ( ) ( )
( ) ( ) k C k
k E k Br k A k
dem
x y
x x
=
+ + = +1
(1)
where
( )
( )
(

=
(
(
(

=
(
(
(
(
(
(
(
(
(
(
(
(

rct AC AC AC
AC
m
m
k
dem
ref
mI
m
AZ
comp
T r f r q p
r
V
I
u
I
I
V
d
I
r
r
), ( , , ,
, ,
2
1
1

y x
(2)
for the azimuth axis. In order to keep the plant order low, it
is assumed that
rdc
= and the measurement is treated as a
state. d is a fictional state intended to represent the constant
bristle friction in steady state. It is modelled as a constant
that subtracts from the motor torque. V
mI
is the component of
motor voltage, V
m
, due to the integral term in the PI current
loop. u
k-1
is a one-step-delayed u signal, which comes from
the output of the rate loop proportional gain. It is necessary
to delay this term in order to manipulate the plant equations
into state space form. The effect of this delay is negligible,
however.
The system has three outputs although only angle, , is to be
controlled and the other two outputs are constrained as
follows:
max
max
150
6 . 4
V V V
I A I
m
m
= <
= <
(3)
3. KALMAN FILTER
The state vector is only partially measurable, hence it is
necessary to estimate the unmeasured states. For this, a
Kalman filter is used. Partition the state vector into measured
and unmeasured states:
| |
T
k dem comp m
u I r
1
= x (4)
| |
T
ref mI m AZ u
I V d I r = x (5)
for the azimuth axis. The relevant matrices must then be
partitioned:
( ) ( ) ( )
( )
( )
( )
( )
( ) k
E
E
k r
B
B
k
k
A A
A A
k
k
k E k Br k A k
dem
u
m
u
m
dem

+
(

+
(

=
(

+
+
+ + = +
2
1
2
1
22 21
12 11
) (
1
1
) ( 1
x
x
x
x
x x
(6)
so that:
( ) ( ) ( )
( ) ( )
( ) ( ) ( ) { }
( ) ( ) 1 1
1 1
1 } 1
1 { 1
1 12
1 11
2 2
21 22
+ =

+ +
+ =
k E k A
k r B k A k
k E k r B
k A k A k
u
dem m m
dem
m u u

x
x x
x x x
(7)
which is in a state-space form. The Kalman filter is then:
( ) ( )
( ) ( ) ( ) { }
( )
( ) ( )
( ) ( ) { } 1 1
1 | 1 1 |
] 2 | 1
1 1 [
2 | 1 1 | 1
2 21
22
12
1 11
+ +
=

+
=
k r B k A
k k A k k
k k A
k r B k A k L
k k k k
dem m
u u
u
dem m m
u u
x
x x
x
x x
x x
(8)
where L is the solution to the Riccati equation.

4. PREDICTIVE CONTROL
Predictive Control is a very powerful controller design
technique, as it minimises a quadratic cost function on the
system output error and control input, whilst also taking into
account constraints on particular signals.
Given a state-space system represented by:
( ) ( ) ( )
( ) ( ) ( ) ( ) k C k k C k
k B k A k
z y
x z x y
u x x
= =
+ = +
,
1
(9)
the aim is to minimise the error between the controlled
output, ( ) k i k | + z , and the reference, ( ) k i k | + r , whilst
restricting the control input change, ( ) k i k | + Au , over a
finite number of steps into the future. If the state is not
available for direct measurement, then the measured output,
( ) k y , is used for state estimation.
The cost function, from Maciejowski (2002), is:
( ) ( ) ( )
( )
( )
( )
( ) ( ) ( )
2 2
0
2
2
|
| |
R Q
U T Z k k k
k i k
k i k k i k k V
u
p
W
H
i
i R
H
H i
i Q
A + =
+ A +
+ + =

=
=
u
r z
(10)
where
( )
( )
( )
( )
( )
( )
( )
( )
( )(
(
(

+ A
A
= A
(
(
(

+
+
=
(
(
(

+
+
=
k H k
k k
k
k H k
k H k
k
k H k
k H k
k
u
p
w
p
w
| 1
|
|
|
,
|
|
u
u
r
r
z
z


U
T Z
(11)
In order to solve the problem, a prediction of the future
controlled outputs, ( ) k i k | + z , is required. In the simplest
case, where the whole state is measured, the expression
below is used:
( )
( )
( )
( ) ( )
( ) k
B A B A
B
k
B A
B
k
A
A
k H k
k k
k
u p p
p p
H H
i
i
H
i
i
H
i
i H
p
U
X
A
(
(
(

(
(
(

+
(
(
(

=
(
(
(

+
+
=

=
0
1
0
1
0
0
1
|
| 1

u x
x
x
(12)
and the controlled output estimate formed with:
( ) ( )
( ) ( ) ( ) ( ) k k k k
k i k C k i k
z
U Z OA + Y + + =
+ = +
1
| |
u x
x z
(13)
If the tracking error is defined as the difference between the
reference vector and the free response of the system:
( ) ( ) ( ) ( ) ( ) 1 Y + + = E k k k k u x T (14)
then ( ) k V may be restated in terms of the input predictions
and tracking error:
( ) ( ) ( ) ( )
( ) ( ) ( )
( )
2
2 2
(

A
E OA
=
A + E OA =
k
k k
k k k k V
U S
U S
U U
R
Q
R Q
(15)
where
Q
T
Q
S S Q = and
R
T
R
S S R = . The aim is to find
( )
opt
k U A to minimise ( ) k V . Ideally, we would have
( ) ( ) ( )
( )
0 =
(

A
E OA
k
k k
U S
U S
R
Q
, equivalently stated as
( )
( )
(

E
= A
(

O
0
k
k
Q
R
Q
S
U
S
S
. This is a least-squares problem
which may be solved for ( ) k U A at each time step, k , but
only ( ) ( ) ( ) 1 | | + A = k k k k k u u u is applied.
Constraints are specified in the following form:
( ) ( ) ( )
0
1
, 0
1
, 0
1
s
(

s
(

s
(

A k
G
k
F
k
E
Z U U
(16)
which may be expressed as a single inequality:
( )
( )
( ) ( ) ( )
( ) s OA
(
(
(

Y + + I

s A
(
(
(

IO
k U
w
g k u k x
f k u F
k
W
F
1
1
1 1
U (17)
On inspection, the overall problem is of the form:
( )
( ) ( ) ( )
( )
( )
( ) ( ) ( )
( ) ( ) ( ) ( )
( )
( ) ( ) ( ) ( ) ( )
( )
( ) ( ) ( )
T T
k
T T T
k
T T
T T
k
k
k k k
k k k k k
k k k k
k k
k
k k
U G U H U
Q G U U H U
Q Q U
U R Q U
U S
U S
U
U
U
R
Q
U
A + A A =
E E + A + A A =

)

E E + E O A
A + O O A
=
(

A
E OA
A
A
A
A
min
min
2
min
min
2
(18)
subject to ( ) s OA k U . This is well known as the
Quadratic Programming (QP) optimisation problem.

5. TRACKING TURRET EXAMPLE
The azimuth and elevation axes are required to follow
realistic reference signals corresponding to target and aircraft
movement. For example, if the aircraft rolls in flight, then
the target appears to move along an arc of a circular
trajectory relative to the turret. Supposing that the turret is at
the origin of a coordinate system, a suitable reference may be
generated from a circle in three-dimensional space.
Begin by specifying the angle, h, in radians, by which the
circle will miss the nadir. If this angle is small, then it will
approximately equal the distance of the centre of the circle
from the origin. If the trajectory is traced out by a unit
vector, then the circle radius is
2
1 h r = .
The remaining parameters are rotation rate, , initial phase
on the circle, , and azimuth angle of the circle, .
Coordinates of the unit vector in an (x,y,z) system may be
computed by transformation. Fig. 2 shows an example when
h=1=0.0175rad and =45.
Fig. 2 Reference Vector Trajectory
The circle is at 45 to the y-z plane in accordance with , and
the origin is approximately 0.0175 from the centre of the
circle, in agreement with h.
The example vector in (x,y,z) is then translated into demands
on the azimuth and elevation of the turret. The azimuth
reference is tan
-1
(y/x) and the elevation reference is sin
-1
(z/1).
Three points on the circle may be used to extrapolate the
azimuth and elevation demands into the future, as depicted in
Fig. 3, corresponding to the darkened arc (Fig. 2).
The reference demand extends 1.5 seconds into the future and
=60/s , =150. Fig. 4 depicts the rate demand in the
example, where azimuth rate peaks at 60rad/s when the
elevation angle is -89. This illustrates the nadir
phenomenon well. If h shrinks to less than 1, the azimuth
rate demand peak will grow, until it becomes infinite when h
is zero. However, due to the motor voltage and current
constraints in equation (3), the maximum acceleration and
velocity of each axis are also constrained. In the azimuth
case, these maxima are 80rad/s
2
and 6.08rad/s respectively.
The predictive controller can cope with the huge rate demand
via the use of future knowledge and constraint handling.
Fig. 3 Angle Reference Trajectory
Fig. 4 Rate Reference Trajectory
In the control example that follows, the above angle reference
will be applied to both the MPC controller and the standard
PI controller.
The sample time must be short in order to accurately estimate
the rapid response of current and voltage in the motor
controller. 1ms is selected, as this is the maximum capability
of the Selex hardware. The output prediction horizon is 150
steps or 0.15 seconds and the input horizon is 10 steps or 0.01
seconds for both azimuth and elevation axes. These are the
maximum horizon sizes without introducing ill-conditioned
matrices and numerical errors. The error weightings are
Q(i)=1+0.075i for azimuth and Q(i)=1+0.15i for elevation.
The input weightings are R(i)=0.4-0.02i in both cases. The
increasing output weights and decreasing input weights
improve the overshoot and settling characteristics, since the
controller gain is effectively increased as the setpoint is
approached. The time constant for the exponential trajectory
from present output to reference trajectory is 0.1 seconds in
both cases.

Figs. 5 and 6 give a comparison between the existing PI
controller and the MPC controller when the above example
reference signal is applied.
Fig. 5 Azimuth Angle
Fig. 6 Elevation Angle
Fig. 5 clearly shows that MPC anticipates the nadir and
begins azimuth rotation earlier. Overshoot is also smaller,
although the PI plot settles whilst the MPC plot exhibits
small limit cycles due to the friction model and estimates.
Fig. 6 demonstrates superior performance by MPC for the
elevation angle, where the offset is negligible compared to
the PI case. At the nadir after 0.5 seconds, the MPC plot
does overshoot, but this is again due to friction, which is now
discussed in terms of state estimates.
The r
AZ
and I
ref
estimates are very accurate and not shown
here. Figs. 7 to 9 show the state estimates of azimuth axis
motor current, disturbance and motor voltage, V
mI
, due to the
PWM integrator.
The flip in sign of the disturbance is responsible for the poor
current and voltage estimates and the limit cycle seen in Fig.
5. A similar effect is seen at the lowest point of elevation,
where the elevation disturbance flips in sign and degrades the
other state estimates.
Fig. 7 Azimuth Current
Fig. 8 Azimuth Disturbance
Fig. 9 Azimuth Voltage from PWM Integrator

Note that the constraint in inequality (3) is obeyed for the
motor current and is also obeyed for motor voltage, but this
cannot be seen directly from Fig. 9. Also, the results are not
greatly affected by plant-model mismatch, as the Kalman
filter tends to express the unmodelled behaviour in terms of
the d state. There is not sufficient space in this paper to
illustrate this fact, however.
6. CONCLUSIONS
A novel application of constrained model predictive control
to a servo-driven tracking turret simulation has been
demonstrated. The turret is modelled in state-space, where
four of the states are measurable and five are estimated using
a Kalman filter. It is demonstrated that around the nadir
point, where the system is driven to saturation limits and will
potentially lose a target, performance of MPC is superior to
that of the existing classical PI controller.
The existence of a reliable reference signal, extrapolated from
the target trajectory, allows MPC to anticipate the large rate
demand near to the nadir. This gives a much earlier response,
which is vital for target acquisition and following in this high
bandwidth tracking application. The system constraints
encapsulated in MPC also produce less overshoot of the
azimuth angle and would do so for the elevation angle, given
better friction estimates.
The main problem with the MPC is that friction in the turret
is not well modelled in state-space, which tends to produce
poor state estimates when friction effects are particularly
active. The friction is modelled as a constant with slight
excitation by white noise, which is acceptable when turret
angles are increasing or decreasing monotonically. However,
when the rate changes sign, the flip in friction sign tends to
give small limit cycles. As further work, it may be possible
to revise the state-space friction model so that better-behaved
estimates are obtained. Indeed, the underlying non-linear
model in the simulation may need revision, as it is unclear
how accurately the real friction of the turret is represented.
Another problem with the MPC is that the 1ms sample time,
dictated by estimating the high bandwidth motor controller,
prevents use of prediction horizons beyond 0.15s due to
numerical errors. This application would benefit from
accurately predicted target trajectories further into the future,
so extended horizons are desirable. There is not much scope
to reduce the sample rate, due to potential aliasing in the
estimator, but more numerically robust QP solvers could be
investigated.
Finally, the MPC controller should be tested with real
hardware, but software limitations in implementing the QP
solver have so far prevented this.
ACKNOWLEDGEMENTS
The authors are grateful to Selex Sensors & Airborne
Systems for their general support, and wish to offer thanks
for the mathematical turret model and Simulink simulation on
which the study was based. Thanks also to the EPSRC for
funding this work on Grant GR/S60396/01. The first author
would like to thank Dr. Nick Brignall and Matt MacDonald
at Selex for their invaluable contribution to this work. Also
thanks to Prof. Mike Grimble, Dr. Reza Katebi and the
University of Strathclyde Industrial Control Centre for
esteemed advice and support.
REFERENCES
Maciejowski, J.M. (2002). Predictive control with
constraints. Prentice-Hall, Pearson Education
Limited, Harlow, England.
Masten, M.K. (1996). Electromechanical systems for optical
target tracking systems. In: Selected Papers on
Precision Stabilization and Tracking Systems
for Acquisition, Pointing, and Control
Applications (Masten, M.K, Stockum, L.A (Ed)),
SPIE Milestone Series, Volume MS 123.
Mitchell, E.E.L., and A.E. Rogers (1968). Quaternion
parameters in the simulation of a spinning rigid body.
In: Simulation: The Dynamic Modeling of
Ideas and Systems with Computers (John
McLeod, P.E. (Ed)).
Robinson, A.C. (1958). On the use of quaternions in
simulation of rigid-body motion, Wright Air
Development Center (WADC) Technical
Report 58-17.

You might also like